1. Exploiting Graphics Processors for High- performance IP Lookup in Software Routers Author: Jin Zhao, Xinya Zhang, Xin Wang, Yangdong Deng, Xiaoming Fu Publisher: IEEE INFOCOM 2011 Presenter: Ye-Zhi Chen Date: 2012/01/11

2. INTRODUCTION  This paper present GALE, a GPU-Accelerated Lookup Engine, to enable high-performance IP lookup in software routers.  Authors leverage the Compute Unified Device Architecture (CUDA) programming model to enable parallel IP lookup on the many-core GPU. 2

3. INTRODUCTION The key innovations include: 1) IP addresses are directly translated into memory addresses using a large direct table on GPU memory that stores all possible routing prefixes 2) Authors also map the route-update operations to CUDA in order to exploit GPU’s vast processing power 3

4. architecture The traditional trie-based routing table in system memory GALE also stores the next-hop information for all the possible IP prefixes in a fixed-size large memory block on GPU, which is referred to as direct table The reason for maintaining two routing tables is that there is an inherent tradeoff between efficiency and overhead when considering a direct table versus trie for particular functionality. 4

5. architecture 5

6. GALE employs a control thread and a pool of working threads to implement the lookup and update requests. The processing of requests is divided into the following steps : 1) Control thread reads a group of requests of the same type from the request queue classified by Lookup, Insertion, Modification, Deletion operations. 2) Control thread activates one idle working thread in the thread pool to process the group of requests. 3) Working threads invoke corresponding GPU and CPU code to perform lookup and update operations 6

7. architecture With GALE, the parallelism in request processing is exploited in two aspects. 1. As the CPU we use also has multi-cores, the working threads can be potentially executed simultaneously. Meanwhile, each working thread will instruct the GPU to launch one kernel. As GTX 470 supports multiple kernels on different SMs, the different request groups thus can be scheduled in parallel 2. Inside one kernel, the group of requests are of the same type. Parallelism is thus achieved by mapping these requests to the CUDA threads, which in turn are scheduled in parallel to different SPs on GPU 7

8. Lookup Authors explore an IP lookup solution with O(1) memory access and computational complexity. each IP address has exactly one-to-one mapping relation with entries in the direct table. They found that over 40% of IP prefixes have the length of 24, and over 99% of the IP prefixes are less than or equal to 24. They leverage this fact and propose a solution that storing all the possible prefixes with length less than 24 in one direct table(which only cost 2 24 =16M memory) and use a separate long table to store the prefixes that are longer than 24 bits 8

9. Lookup Only one address translation is enough to find the corresponding routing entry. Meanwhile, only one memory access is required to fetch the next-hop destination. 9

10. UPdate route-update will involve operations to both the direct table and the trie structure. As direct table has a fixed size, we actually need not allocate or release entry space upon route-update operations. To alleviate the trie-traversal operations in GPU, we introduce a length table to denote the prefix length. 10

11. UPdate 1) Insertion and Modification: Insertion and modification are essentially the same operation for direct table 11

12. UPdate 2) Deletion : The deletion operation is similar to modification except that during updating the range of entries, the next-hop information is modified to the parent node’s nexthop information in the trie Deleting an entry is as follows: replacing the nexthop and prefix length of the updated entry with the parent’s nexthop and prefix length. The parent node is obtained from the trie-traversal during deleting the entry node in trie by CPU. 12

13. UPdate 13

14. The routing dataset used in the experiments are from FUNET and RIS Desktop PC equipment : ($1122) Core i5 750 CPU 2.66Ghz (with 4 cores) NVIDIA GeForce GTX470 with 1280MB of global memory and 448 stream processors ($428) OS: 64-bit x86 Fedora 13 Linux distribution with unmodified kernel 2.6.33 CUDA version : 3.1 14 Experiment

15. the experiments use the following default configuration: With 8 working threads. Every CUDA thread (running on GPU ) takes just one lookup request, or updates one entry while performing update tasks. Every CUDA grid contains 44 thread blocks, and every block contains 256 1 threads. 15

16. Experiment 16

17. Experiment 17

18. Experiment 18

19. Experiment 19